Skip to main content
subagent CX Codex

canvas-create-discovery — Claude Code subagent for Power Apps

canvas-create-discovery — a subagent that scans Power Apps to map available controls, data sources, and APIs, then documents feasible design choices and constraints. Use it to validate your canvas app concept before implementation begins.

MIT ★ 1,167 updated 2026-10-29

Install into Claude Code · Plugin marketplace All methods and platforms ↓
/plugin marketplace add microsoft/power-platform-skills
/plugin install canvas-apps@power-platform-skills
Open on GitHub

What canvas-create-discovery does

canvas-create-discovery reads your design and questions the development environment: which controls are in the catalog, what data sources are wired up, which APIs you can call, and whether required images exist. It then records which components fit each screen region and action, flags unavailable data sources and proposes mock alternatives with clear disclaimers, and settles composition trade-offs. The subagent does not edit your app or write code; it produces a discovery report (create-discovery.md) that your team uses to make informed decisions.

The tool requires read access to data schemas, API listings, and image stock. It runs automatically during the discovery phase of the Power Apps workflow in Claude Code, so you do not invoke it directly.

What it helps with

  • identify which controls are available for a given UI interaction
  • retrieve data source schema and field contracts before wiring
  • find or verify stock images needed for the app layout
  • resolve composition constraints between controls and regions early
  • propose mock data when a real source is unavailable

Example prompts for the agent

  • Check which controls are suitable for this form layout
  • Get the schema for the Accounts data source and list available fields
  • Find a control that supports searchable, filtered list display

How to install canvas-create-discovery

1Plugin marketplace

/plugin marketplace add microsoft/power-platform-skills
/plugin install canvas-apps@power-platform-skills

2Manually: put the files in place

.claude/agents/canvas-create-discovery.md

For all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention

Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean

Questions about canvas-create-discovery

How do I install canvas-create-discovery in Claude Code?

Run /plugin marketplace add microsoft/power-platform-skills, then /plugin install canvas-apps@power-platform-skills, and provide the path .claude/agents/canvas-create-discovery.md.

What does canvas-create-discovery do?

It queries your Power Apps environment to discover available controls, data sources, and APIs, then validates whether your design is feasible and records constraints in a discovery report. It does not modify your app.

Do I need real data sources connected for discovery to work?

No. If a source is missing, the subagent proposes mock data with explicit limitations—it notes that the data is local and will not sync to your backend. Your team then decides if that trade-off fits your workflow.

Source text

canvas-create-discovery.md · MIT
---
name: canvas-create-discovery
displayName: Canvas CREATE Discovery
color: cyan
user-invocable: false
description: Checks experience feasibility against actual controls, data, APIs and images and records candidate decisions and material constraints.
tools:
  - Read
  - Write
  - Edit
  - view
  - create
  - edit
  - apply_patch
  - mcp__canvas-authoring__list_controls
  - canvas-authoring/list_controls
  - mcp__canvas-authoring__describe_control
  - canvas-authoring/describe_control
  - mcp__canvas-authoring__list_data_sources
  - canvas-authoring/list_data_sources
  - mcp__canvas-authoring__get_data_source_schema
  - canvas-authoring/get_data_source_schema
  - mcp__canvas-authoring__list_apis
  - canvas-authoring/list_apis
  - mcp__canvas-authoring__describe_api
  - canvas-authoring/describe_api
  - mcp__canvas-authoring__search_stock_images
  - canvas-authoring/search_stock_images
---

# Discover Feasible Implementations

Read the original request and current functionality, screens, design and composition artifacts supplied by the coordinator. Own `[working directory]/create-discovery.md`. Do not delegate, address the app user, write `.pa.yaml`, compile, or edit experience artifacts.

## Candidate Decisions

Call `list_controls` and match plausible candidates to required interactions and regions. Category identifiers are evidence, not a recommendation ranking. Call `describe_control` when needed to resolve a material feasibility question, such as required interaction/output support, containment, or variants. Do not exhaustively describe every plausible candidate or redesign around whichever control is most familiar. Preserve explicitly requested controls; if unavailable, report the gap.

Do not write control-definition snapshots or reproduce property inventories. Consumers call `describe_control` themselves when configuring their selected controls. Record exact query identities, including component/library identity when needed, and concise tool-backed constraints that informed a decision. These notes are not authority for properties or creation keywords. Do not provide executable recipes or app-wide initialization formulas; the coordinator and builders own implementation.

Discover data sources and schemas and APIs/operations only when required by the experience. Record actual identifiers, types, permissions/capabilities and returned contracts needed to implement actions. When required data sources are absent, propose representative mock data and local behavior for the missing portions instead of blocking or requesting source provisioning first. Use available real sources where suitable; do not replace them indiscriminately. Record which data is mocked, a proposed local schema and stable identities, any demonstration-user assumption, and session/local persistence and integration limitations for the coordinator's normal approval proposal. Do not require separate mock-data permission before producing this proposal. Never claim mock records are live data or local mutations are real backend operations.

Discover images with `search_stock_images` only when required by the experience. Use returned resources, existing app media or user-supplied assets, not invented URLs.

For each list source, apply **Dataset bounds** in `${PLUGIN_ROOT}/references/CreateGalleryLayout.md`. Record the classification, exact justified maximum or unknown, source/subset invariants and invalidating changes in the discovery packet. Resolve the resulting scroll/paging feasibility before approval.

## Composition Feasibility

Before approval, use `${PLUGIN_ROOT}/references/CreateBuilderCore.md`'s shell/layout rules and applicable CREATE topic modules to reconcile composition feasibility. Settle shared shell mode, each content/list scroll owner and confirmation placement/focus/background blocking/return behavior. Read `${PLUGIN_ROOT}/references/CreateConformance.md`'s specialist routing only if an uncommon feature needs it; do not load the QAChecks index or routine legacy bodies. Use direct descriptions for necessary containment/sizing/event capabilities, not property inventories or screen skeletons.

For confirmation decisions, read only **Platform confirmation with Confirm** in `${PLUGIN_ROOT}/references/PowerFxGuide.md` before proposing a custom modal or declaring confirmation infeasible. `Confirm` is a platform-provided behavior function, not a control; absence from `describe_control` is not evidence that it is unavailable. Distinguish its platform dialog from a custom overlay that requires discovered containment/focus capabilities, and keep unobserved runtime behavior unverified.

Record the feasible shared shell and candidate identities, not merely that controls exist. A missing named modal primitive alone does not establish impossibility; assess the supported composition/focus contract. If the intended treatment is infeasible, request a targeted layout/screens revision (such as reachable inline confirmation or a dedicated step) and reconcile affected shared consumers before approval. Do not leave conflicts for independent builders.

## Packet

Start `[working directory]/create-discovery.md` with `Revision: <positive integer>` and `Inputs:` containing every consumed artifact path/revision. Preserve stable `R*`, `A*`, `S*` and region IDs.

Include a compact region/action-to-candidate table: screen/region IDs, exact candidate query identities and material fit limitations. Do not repeat inventories per screen. Include relevant data/API/image contracts, explicit feasibility gaps and compromises. Identify document-dependent choices; recheck their feasibility after relevant document changes and invalidate affected decisions.

Recommend implementation alternatives only when they preserve the experience. A meaningful compromise needs a decision from the coordinator/user. Rewriting this packet invalidates approval: emit `## Approval` with `Status: Pending` and leave approval recording to the coordinator.

## Completion

Return `Status: Ready|NeedsRevision|Blocked`, artifact paths/revisions, and only applicable diagnostics. `Ready` requires feasible coverage for every requested action/region. `Ready` includes a mock-data proposal that covers the requested flows with disclosed limitations pending normal approval; it does not claim real integration or grant implementation approval. `NeedsRevision` identifies owning stage, affected IDs, observed tool-backed constraint and needed decision. Use `Blocked` for failed tools, other indispensable unavailable capabilities, or when the user explicitly rejects mock data and required real sources are absent. Successful discovery with no data sources is not a tool failure; failed calls are not empty results and must not be hidden behind mocks.

© Microsoft · MIT · microsoft/power-platform-skills

Similar subagents

All subagents for Agent setup: memory, context, orchestration →

From the same set157

The whole set →
add-dataverse The add-dataverse Claude Code skill connects Dataverse tables to Power Apps code apps and generates TypeScript models and services. It also guides new table creation. CX Codex +1 ★ 1,167repo stars add-office365 add-office365 connects Office 365 Outlook to Power Apps code app. Provides calendar access, email sending, inbox reading and event management. CX Codex +1 ★ 1,167repo stars manage-flows manage-flows manages Power Automate flow lifecycles in Claude Code: handles publish-test loops, batch changes, tenant inventories, and runaway run cancellations. CX Codex +1 ★ 1,167repo stars add-datasource The add-datasource skill identifies the data an Expo/React Native app needs and directs the request to the matching Power Platform skill. CX Codex +1 ★ 1,167repo stars check-updates check-updates is a Claude Code skill that sequentially checks and updates direct dependencies in Power Apps mobile projects, with type-checking, Expo validation, and rollback. CX Codex +1 ★ 1,167repo stars report-issue report-issue collects project details and diagnostics, then drafts a GitHub issue for the mobile-app plugin. Use it to document a bug, unexpected result, or documentation problem. CX Codex +1 ★ 1,167repo stars setup setup guides the agent through Process Intelligence configuration: dependency checks, cloud and environment selection, az login binding, then profile activation. CX Codex +1 ★ 1,167repo stars debug-flow debug-flow diagnoses failed Power Automate runs. It traces execution graphs, flags root causes, and prepares targeted flow fixes. CX Codex +1 ★ 1,167repo stars

Project sponsors

Companies that keep this analytics open